Symplectic Runge-Kutta methods satisfying effective order conditions

Transcription

1 /30 Symplectic Runge-Kutta methods satisfying effective order conditions Gulshad Imran John Butcher (supervisor) University of Auckland 5 July, 20 International Conference on Scientific Computation and Differential equations (SciCADE 20) University of Toronto, Canada

2 2/30 Introduction The idea of effective order was first introduced by Butcher in 969 for explicit Runge Kutta methods as a mean of overcoming the 5 th order 5 stage barrier. 2 This was extended to Singly Implicit Runge Kutta methods by Butcher and Chartier in The idea was later used for Diagonally Extended Singly Implicit Runge Kutta methods by Butcher and Diamantakis in 998 and by Butcher and Chen in The accuracy of symplectic integrators for Hamiltonian systems was enhanced using effective order by M A Lopéz-Marcos, J M Sanz-Serna and R D Skeel in 996.

3 Introduction The idea of effective order was first introduced by Butcher in 969 for explicit Runge Kutta methods as a mean of overcoming the 5 th order 5 stage barrier. 2 This was extended to Singly Implicit Runge Kutta methods by Butcher and Chartier in The idea was later used for Diagonally Extended Singly Implicit Runge Kutta methods by Butcher and Diamantakis in 998 and by Butcher and Chen in The accuracy of symplectic integrators for Hamiltonian systems was enhanced using effective order by M A Lopéz-Marcos, J M Sanz-Serna and R D Skeel in 996. The purpose of this presentation is to develop some special symplectic effective order methods with low implementation costs. 2/30

4 3/30 Differential Equations with Invariants Consider an initial value problem y (x) = f(y(x)), y(x 0 ) = y 0. () Suppose Q(y) = y T My is a quadratic invariant, that is or Q (y)f(y) = 0, y T Mf(y) = 0. Methods which conserve quadratic invariants are said to be Canonical or Symplectic.

5 4/30 Outline Effective order Algebraic interpretation Computational interpretation

6 4/30 Outline Effective order Algebraic interpretation Computational interpretation 2 Symplectic Runge Kutta methods Superfluous and Non superfluous trees Order conditions for Symplectic methods

7 4/30 Outline Effective order Algebraic interpretation Computational interpretation 2 Symplectic Runge Kutta methods Superfluous and Non superfluous trees Order conditions for Symplectic methods 3 Effective order with symplectic integrator New Implicit Methods Cheap implementation Transformation Starting method

8 4/30 Outline Effective order Algebraic interpretation Computational interpretation 2 Symplectic Runge Kutta methods Superfluous and Non superfluous trees Order conditions for Symplectic methods 3 Effective order with symplectic integrator New Implicit Methods Cheap implementation Transformation Starting method 4 Numerical Experiments

9 4/30 Outline Effective order Algebraic interpretation Computational interpretation 2 Symplectic Runge Kutta methods Superfluous and Non superfluous trees Order conditions for Symplectic methods 3 Effective order with symplectic integrator New Implicit Methods Cheap implementation Transformation Starting method 4 Numerical Experiments 5 Conclusions

10 Outline Effective order Algebraic interpretation Computational interpretation 2 Symplectic Runge Kutta methods Superfluous and Non superfluous trees Order conditions for Symplectic methods 3 Effective order with symplectic integrator New Implicit Methods Cheap implementation Transformation Starting method 4 Numerical Experiments 5 Conclusions 6 Future work 4/30

11 5/30 Effective order Algebraic interpretation Algebraic interpretation We introduce an algebraic system which represents indvidual Runge Kutta methods and also composition of methods. 2 This centres on the meaning of order for Runge Kutta methods and leads to the possible generalisation to the effective order. 3 We introduce a group G whose elements are mapping from T (rooted trees) to real numbers and where the group operation is defined according to the algebraic theory of Runge Kutta methods or to the theory of B series. 4 Members of G represents Runge Kutta methods with E representing the exact solution. That is, E : T R is defined by E(t) = γ(t), t

12 Effective order Algebraic interpretation Table: Group Operation t r(t) α(t) β(t) (αβ)(t) E(t) α β α β 0 +β 2 α 2 β 2 α 2 β 0 +β 2 +α β 2 3 α 3 β 3 α 3 β 0 +β 3 +α 2 β +2α β α 4 β 4 α 4 β 0 +β 4 +α β 2 +α 2 β 4 α 5 β 5 α 5 β 0 +β 5 +3α β 3 +α 2 β +α 3 β 4 α 6 β 6 α 6 β 0 +β 6 +α β 4 +α β α 2 β 2 +α 2 β 2 +α α 2 β α 7 β 7 α 7 β 0 +β 7 +2α β 4 +α 3 β +α 2 β α 8 β 8 α 8 β 0 +β 8 +α 2 β 2 +α β 4 +α 4 β 24 6/30

13 7/30 We introduce N p as a normal subgroup, which is defined by N p = {α G : α(t) = 0,whenever r(t) p} A Runge- Kutta method with group element α is of order p, if it is in the same coset as EN p, that is αn p = EN p A Runge- Kutta method has an effective order p if there exist another Runge - Kutta method with corresponding group element β, such that βαn p = EβN p

14 8/30 Computational interpretation Computational interpretation The conjugacy concept in group theory provides a computational interpretation of the effective order. This means that, every input value for effective order method α is perturbed by a method β. Therefore the starting method β offers some freedom of the order conditions of the effective order method α. Every output value could also be perturbed back to the origional trajectory using method β. α n β β x 0 E n x n

15 9/30 Symplectic Runge Kutta methods Symplectic Runge Kutta methods A Runge Kutta method is said to be canonical or symplectic if the numerical solution y n also has the quadratic invariant Q(y n ) i.e. y n,y n = y n,y n A method has this property if and only if, for all i,j. b i a ij +b j a ji b i b j = 0

16 Symplectic Runge Kutta methods Superfluous and Non superfluous trees Superfluous and Non superfluous trees It is a consequence of the symplectic condition that if τ and τ 2 are rooted trees corressponding to the same tree τ then φ(τ ) = γ(τ ), φ(τ 2) = γ(τ 2 ) For Symplectic Runge Kutta methods, we distinguish trees in two ways. Superfluous trees, Non Superfluous trees. 2 τ τ τ 2 0/30

17 /30 Symplectic Runge Kutta methods Superfluous and Non superfluous trees Order conditions corresponding to non superfluous trees are transformed into one order condition. a τ b τ a τ b

18 2/30 Symplectic Runge Kutta methods Order conditions for Symplectic methods Order Conditions for Symplectic Runge Kutta methods The number of order conditions for symplectic Runge Kutta methods is less than the number of order conditions for a general Runge Kutta method. Case : Suppose the method is of order at least, ( b i = ), i b i a ij + i,j i,j b j a ji i,j b i b j = 0 i,j b i a ij = 2 Therefore for symplectic Runge Kutta method the second order condition is automatically satisfied and hence not required.

19 Symplectic Runge Kutta methods Order conditions for Symplectic methods Order Conditions for Symplectic Runge Kutta methods Case 2 : Suppose the method is of order at least 2, ( i b i c i = ), Consider the symplectic condition, 2 b i a ij +b j a ji b i b j = 0 Multiply with c j and take summation, b i a ij c j + b j c j a ji b i b j c j = 0 i,j i,j i,j ( b i a ij c j 6 )+( b j cj 2 3 ) = 0 i,j j

20 3/30 Symplectic Runge Kutta methods Order conditions for Symplectic methods Order Conditions for Symplectic Runge Kutta methods Case 2 : Suppose the method is of order at least 2, ( i b i c i = ), Consider the symplectic condition, 2 b i a ij +b j a ji b i b j = 0 Multiply with c j and take summation, b i a ij c j + b j c j a ji b i b j c j = 0 i,j i,j i,j ( b i a ij c j 6 )+( b j cj 2 3 ) = 0 i,j j

21 4/30 Symplectic Runge Kutta methods Order conditions for Symplectic methods Order General RK method Symplectic RK method Table: Order conditions for general and symplectic Runge Kutta methods up to order 4.

22 Symplectic Runge Kutta methods Order conditions for Symplectic methods Gauss method For example, consider applying Gauss method to the harmonic oscillator problem given below: The energy is given by, q = p, p = q. H = p2 2 + q2 2. The exact solution is, [ ] [ ][ ] p(x) cos(x) sin(x) p(0) =. q(x) sin(x) cos(x) q(0) This problem describes the motion of a unit mass attached to a spring with momentum p and position co-ordinates q defines a Hamiltonian system. 5/30

23 6/30 Symplectic Runge Kutta methods Order conditions for Symplectic methods we consider the two stage order four Gauss method, (2) 2 2 Gauss method approximately conserve the total energy of the himiltonian system.

24 7/30 Effective order with symplectic integrator Effective order with symplectic integrator We show a way of analyzing methods of effective order 4 having symplectic condition with three stages. (βα)( ) = β( )+α( ) (βα)( ) = β( )+2β( )α( )+β 2 ( )α( )+α( ) (βα)( ) = α( )+3β( )α( )+3β 2 ( )α( ) +β 3 ( )α( )+β( )

25 Effective order with symplectic integrator New Implicit Methods New Implicit methods We present two examples of symplectic effective order methods: Method A - has real and distinct eigenvalues Method B - has complex eigenvalues /30

26 9/30 Effective order with symplectic integrator Cheap implementation Cheap implementation Since method A have real eigenvalues, it is therefore of interest.this is because we can obtain a cheaper implementation. Here we are considering only method A. The general form of an s-stage implicit method is y n+ = y n +h Y i = y n +h s b i f(x n +hc i,y i ), i= s a ij f(x n +hc j,y j ) j= The stage equations can be written in the form Y = e y n +h(a I m )F(Y)

27 20/30 Effective order with symplectic integrator Cheap implementation We use modified Newton Raphson iteration scheme to solve the above equation. This can be defined as where M Y [k] = G(Y [k] ) Y [k+] = Y [k] + Y [k] M = I s I m h(a J) G(Y [k] ) = Y [k] +e y n +h(a I m )F(Y) The total computational cost in this scheme include the evaluation of F and G, the evaluation of J, the evaluation of M, LU factorization of the iteration matrix, M, back substitution to get the Newton update vector, Y [k].

28 Effective order with symplectic integrator Transformation Transformation To reduce computational cost of fully implicit RK method, we use transformation. The transformation matrix T for method A is given by T = The transformation matrix has the property, T AT = where , , and are three distinct real eigenvalues 2/30

29 22/30 Effective order with symplectic integrator Starting method Starting method Solution of these equations give the starting method (α)( ) = (α)( ) = 2β( )+ 3 (α)( ) = 3β( )+3β( )+ 4 which is given by

30 23/30 Numerical Experiments Numerical Experiments The Kepler s problem x = y, x 2 = y 2, x y =, y (x 2 2 = +x2 2 )3 2 (x 2 +x2 2 )3 2 x 2 where (x,x 2 ) are the position coordinates and (y,y 2 ) are the velocity components of the body. (x,x 2,y,y 2 ) = ( e,0,0, (+e)/( e)), H = 2 (y2 +y2) 2. x 2 +x2 2

31 24/30 Numerical Experiments Kepler problem (e=0,h=0.0, n = 0 6 ) Graph for Hamiltonian: error VS time 20 x

32 25/30 Numerical Experiments Kepler problem (e=0.5,h=0.0, n = 0 6 ) For Hamiltonian: error VS time 20 x

33 26/30 Numerical Experiments The simple Pendulum p = sin(q), q = p, (p,q) = (0,2.3). H = p2 2 cos(q).

34 27/30 Numerical Experiments Simple Pendulum (h=0.05, n = 0 6 ) For Hamiltonian: error VS time 6 x x 0 4

35 28/30 Conclusions Conclusions For problems that conserve some sort of invariant structure, it is a good idea to use numerical methods which mimic this behaviour. 2 Symplectic Runge Kutta methods have this role for many important problems. 3 Because of greater flexibilty,effective order methods can provide greater efficiency as compared with methods with classical order. 4 It is possible to obtain cheap implementation cost if A has real eigenvalues. 5 These methods are suited for parallel computers which have very large number of processors.

36 29/30 Future work Future work Error estimates 2 Working on implicit methods with optimal choices of parameters. 3 Construct general linear methods with closely related properties. 4 Generalization of effective order on partioned Runge Kutta methods for separable Hamiltonian.

37 30/30 Future work THANK YOU